bioRxiv · 10.64898/2026.11.10.765812
Transfer learning for identification of filamentous fungi using culture and microscopy images.
Abstract
Culture morphology and microscopy images are required for identification of moulds. Automated image classification is achievable in specific domains via transfer learning of pre-trained convolutional neural networks (CNNs). The study aimed to evaluate the classification accuracy of ResNet50 CNN, for image-based classification of seven commonly isolated filamentous fungi. A total of 15,400 images were collected from seven filamentous fungi. For each species, 1,100 images each were collected for two image modalities: culture front, and microscopy, resulting in 2,200 images per class. Images from each modality for each fungus, were divided into training, validation, and test sets. The ResNet-50 model was trained and validated using four approaches: culture images (n=7,000), microscopy images (n=7,000), ensemble approach incorporating both culture and microscopy images (n=14,000), and combined culture and microscopy grid images (n=7,000). Performance was evaluated using the independent test dataset comprising images from previously unseen fungal strains, which included 700 images each, for culture, microscopy and grid images, and 1,400 images for assessment of ensemble training approach. We applied transfer learning, using both ResNet50 with and without a hierarchical approach, and evaluated their performance on the test image sets. The highest classification accuracy was achieved using the culture and grid image models, both of which achieved overall accuracy of 100%. The microscopy only and ensemble models achieved overall accuracies of 99.4% and 98.8%, respectively.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Hali Mathew, S., Davis, J., Davis, R.. 2026-11-16. Transfer learning for identification of filamentous fungi using culture and microscopy images.. https://doi.org/10.64898/2026.11.10.765812
Cite the original work for its findings. Save a collection to share your selection of sources.